Object Detection and Counting in UAV Images Based on Deep Learning

Object detection in UAV images is an important part of intelligent traffic control. However, there are many problems such as diverse target scales, small targets and occluded targets in UAV images. To solve these problems, an improved network structure of YOLOv5s is proposed. Firstly, a decoupling detection module is introduced in the head of the network to improve the detection accuracy and speed up the convergence. Secondly, a small object detection layer was added to the original network structure and multi-scale feature fusion was performed to enhance the prediction ability of the algorithm for multi-scale objects and small objects. Finally, a pedestrian detection and counting model of aerial images based on deep convolutional network is designed to count the current crowd and vehicle traffic in the form of quantity detection. Our proposed algorithm model’s mean average precision (mAP0.5) in the publicly available VisDrone2019 dataset attains 40.3% while fulfilling the requirement of real-time detection, which is 5.9% higher than the YOLOv5 algorithm. The improved model’s detection accuracy surpasses those of other mainstream algorithms by a considerable extent.